Wednesday , May 5 2021

AI can now rewrite familiar images with a 3D printer

Can artificial intelligence and 3D printer help fight the greatest threat to the art world?

A team from the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) finds it so.

The new system that aims to design image reproductions uses 3D printing and deep learning to "authentically recreate your favorite images" – regardless of the lighting conditions or placement.

"If you simply reproduce the color of the image as it appears in the gallery, it may look different in your home," says co-author of the study, Changil Kim, Ph.D. in CSAIL, in a statement.

Thus, instead of using the traditional four fixed inks (cyan, magenta, yellow, black) found in 2D printers, researchers applied a special technique called "color matching".

The team trained a deep learning model to predict the optimal bunch of different inks (through MIT CSAIL)

The process includes a 3D printer and 10 different transparent inks arranged in very thin layers – similar to trays and chocolate from the whale floor bar.

By combining this method with a decade-long semantic approach (which creates a gradient-like effect across points), they were able to better capture "shades of colors".

Based on experimental replications of various oil paintings created by a collaborator of artists, the team has discovered that RePaint is more than four times more precise in producing accurate shades of colors from state-of-the-art physical models.

"Our system works under any lighting conditions, which shows a much greater ability to reproduce colors than almost any previous work," said Kim.

However, there is one catch: CSAIL's facsimile is just the size of the business card. After all, 3D printing is not cheap.

In the future, however, they expect that advanced commercial printers will provide larger images, paving the way for a more efficient system.

However, the question remains: Which inks should be used for which images?

As with many tasks today, people have carried the burden of selecting a deep learning model that can predict the optimal bunch of shades in each canvas surface.

RePaint's creators predict its use in the processing of home-based artworks, the protection of originals in museums, and the creation of prints of historical works.

The researchers found that RePaint is more than four times more precise than the most sophisticated physical models in creating accurate shades of colors for different artworks (through MIT CSAIL)

However, the program has some ways to go before it can duplicate the duplicate of "Stellar Night".

For starters, certain colors can not be fully reproduced, such as cobalt blue, due to the limited ink library, which machine engineer Mike Foshi hopes to expand soon.

And, as you can see in the images and videos above, there is something particularly missing in the imitation of RePaint: texture. The team will continue to work on better details, eventually hoping to create special effects like glossy and matte finish.

"The value of fine arts has grown rapidly in recent years, so there's a growing tendency to be closed in warehouses, far from the public eye," Fossy said. "We are building the technology to change this trend and create cheap and accurate reproductions that everyone can enjoy."

A full report was published online this week by MIT CSAIL.

More coverage at

Source link